MF-TLMLP: A multi-channel fusion-based multi-fidelity transfer learning method for predicting the cutting performance of conical picks

Conical picks are critical components in key excavation equipment (including shearers, continuous miners, and roadheaders) employed for rock fragmentation in mining and civil engineering. This study proposes the multi-fidelity transfer learning multi-layer perceptron (MF-TLMLP) method to predict cutting performance. Addressing the scarcity of high-fidelity (HF) data, MF-TLMLP integrates low-fidelity (LF) and HF data through a multi-channel fusion and transfer learning approach. Simulations using Latin hypercube sampling (LHS) generated 2500 LF samples and 500 HF samples. Compared to traditional Multi-Layer Perceptron (MLP) and transfer learning models, MF-TLMLP demonstrated superior performance, achieving a higher coefficient of determination (R ² ) and a lower mean squared error (MSE). Morris screening based on mean absolute elementary effects revealed output-dependent sensitivities, with cutting thickness dominating the force responses and cutting speed dominating the pulse-number responses within the sampled domain. These results demonstrate the potential of MF-TLMLP for predicting HF numerical cutting responses for the investigated rock properties, parameter ranges, and adjacent double-pick configuration.